AI Technology Is Helping Transform Unreadable 3,000-Year-Old ‘Chunks of Charcoal’ into Rediscovered Ancient Texts

Nearly 300 years ago, archaeologists in ancient Herculaneum, once a wealthy Roman town, uncovered 1,785 papyrus scrolls in a residential complex that has since become known as the Villa of the Papyri. Situated near Pompeii, just 11 miles from the base of Mount Vesuvius, Herculaneum was home to thousands of elite Roman citizens seeking a coastal retreat. While the site, with its furniture, rich frescoes, intact upper floors, and original wooden balconies, was better preserved than Pompeii, the scrolls—which were discovered in 1752—remained illegible hunks of carbonized ash.

That is, until now, due to the rapid evolution of artificial intelligence technologies.

Though AI has more recently entered the zeitgeist, the Herculaneum breakthrough goes back nearly 20 years, to when computer scientist Brent Seales first became fixated on what to do with “a damaged book you can’t even open.”

As Seales told ARTnews, this question— “Does it get left behind?”—ultimately drove him to virtually map and make legible the surfaces of the ancient scrolls.

The Villa of the Papyri, named for its massive library of scrolls, was owned by Julius Caesar’s father-in-law, Lucius Calpurnius Piso Caesoninus. The scrolls, which represent the only known library of its kind and size from the classical world, are thought to contain significant philosophical and literary texts by preeminent ancient Greek and Roman scholars. Since their discovery, they have been placed in major institutions, including Oxford’s Bodleian Library, the Institut deFrance, and the Getty Museum.

In 2005, Seales, a professor at the University of Kentucky, Lexington, heard about institutions beginning to digitize and digitally restore notable texts in their collections. As an expert in machine learning, AI, and computer vision, he began to wonder if he might be able to use those tools to “virtually unwrap” the Herculaneum scrolls.

After four years of research, Seales had a working prototype and persuaded the Institut de France to provide him the first scroll, on the condition that he set up his system on-site. That initial try, however, was limited by the technology available at the time, as the resolution wasn’t “super high” and the “size of the data” exceeded their computing power.

“We didn’t have a computer when we got back from our first scanning session that could actually load the whole thing into memory,” Seales said, explaining that the data was so massive that the images of the scroll couldn’t be viewed all at once. “Now we have all of this cloud computing, and you can throw things in a system that has massive amounts of memory.”

Despite early strides, the technology was still not capable of completing the “virtual unwrapping” that Seales envisioned. By 2015, however, cameras had been developed that could capture tomography—specialized X-ray imaging—at a high spatial resolution, leading experts to conclude that the Herculaneum papyri still contained writing.

On another project, a team led by Seales was able to successfully trial a new technique, using X-ray tomography and computer vision, on the En-Gedi Scroll, which had been discovered in a region west of the Dead Sea in Israel. Without opening it, they found that the scroll contained writing from the book of Leviticus—a notable text among Judeo-Christians.

While Seales was able to confirm that the Herculaneum scroll contained text via the same “virtual unwrapping” method, the ink used was made from carbon, which is not chemically different enough to decipher from the burned papyri; the density of the ink and the papyri proved to be too similar to be read using the X-ray technology. This observation led to the decision to use AI as an intermediary to enhance and make legible the ink, a pursuit that resulted in the establishment of a competition known as the Vesuvius Challenge.

(It should be noted that, in addition to technological advancements, improvements have also been made in conservation methodology—a change that has impacted collections’ willingness to allow experimental processes to be conducted on priceless artworks and artifacts in projects such as the Rijksmuseum’s Operation Night Watch. When Seales initially carried out his experiments, it was far more difficult to find institutions willing to have their holdings scanned.)

In fall 2022, Nat Friedman, the former CEO of GitHub, heard about Seales’ work and proposed an open contest to push along the research. Seales was initially hesitant, he said, but after failing to raise research funds, he agreed. Friedman and entrepreneur Daniel Gross, with whom he’d been investing in the AI space, put up $125,000 to launch the Vesuvius Challenge. An additional $1 million was raised from other Silicon Valley investors and social media users. For the Challenge, Seales shared his software and high-resolution scans with participants who were asked to create machine-learning models that could distinguish the text from the carbonized ash.

Since that time, the Vesuvius Challenge has awarded tranches of prize money to participants at different stages of the contest, with each round having a particular target goal. In one round that ended in July 2023, a $100,000 award was disbursed to 10 winners who competed among several thousand participants to see who could improve the AI method for detecting the ink in the scrolls. Last year’s grand prize of $850,000 was set for reading four passages of at least 140 characters each before the end of 2023.

The research, code, and methods for each round were released to participants so they could build on each other’s work.

The Challenge saw a breakthrough last October, when US physicist and entrepreneur Casey Handmer noticed a texture like cracked mud in the scans that formed Greek letters. Luke Farritor, an undergraduate computer science student at the University of Nebraska-Lincoln, then used this texture to develop a machine-learning algorithm that identified the word porphyras (purple)—a discovery for which he won an individual prize. Berlin-based PhD student Youssef Nader then developed clearer images of the text.

In the end, 18 submissions were received. After a jury reviewed the code, 12 submissions were presented to a committee of papyrologists, who assessed legibility and transcribed the text. Only one team—formed by Farritor, Nader, and Swiss robotics student at the Swiss Federal Institute of Technology Zurich Julian Schilliger—met the criteria of reading the four passages with at least 85 percent readability. For their efforts, they split a prize of $700,000.

The translated text revealed a previously unknown philosophical work on the senses and pleasure, discussing music, the taste of capers, and the color purple, along with a possible description of known flautist Xenophantus, who had been mentioned in texts by ancient authors Seneca and Plutarch.

While none of the members of the team could read ancient Greek, Nader told ARTnews that he became intimately familiar with the writing hidden in the scrolls. “The process required me to trace the writing to create black and white images of the ink, [akin to] tracing,” he said. “I came to learn how this specific scribe used to write his letters, and how he would draw a letter from a certain point. And you can see, with high resolution [scans], ink deposits where he starts drawing the letter and how he used to draw it.”

For Domenico Camardo,an archaeologist at the Herculaneum Conservation Project, the Vesuvius Challenge has been nothing short of ground-breaking.

“Knowing the enormous effort for reading the Herculaneum papyri that the papyrologists working at the Herculaneum Papyri workshop of the National Library of Naples have been making for decades, I was amazed by how AI ​​managed, without unrolling and therefore risking destroying the charred scrolls, to recognize letters, then words, until entire sentences are reconstructed,” Camardo told ARTnews in an email.

A new iteration of the Vesuvius Challenge has been announced for 2024, with the goal of deciphering at least 90 percent of the four scanned scrolls by the end of the year. The ultimate goal, according to Seales, is to continue improving the speed and accuracy of the technology while also increasing the number of scrolls read.

“Every little bit we learn has the potential to move the needle,” he said. “You honestly don’t know when the really big discovery will happen. The exciting thing about Herculaneum is [that] every one of those books could be anything.”

While any number of these scrolls could contain an impactful work of previously unknown literature or history, it is important to note that they are only one part of a larger library that has yet to be discovered. Seales hopes these continued developments will spur further investigations at Herculaneum, as entire sections have yet to be excavated, and a primary library has not been identified. As such, there lies the potential for thousands more scrolls to be found buried underneath the ash.

The new techniques discovered via the Vesuvius Challenge can be applied to the deciphering of other texts as well—an impact that has not gone unnoticed.

“I’ve seen some promising results from Egyptian scrolls that I also got access to from a project here in Berlin,” said Nader. Through the Herculaneum models, he explained, “they have learned something about papyrus and ink, and they work sometimes for Egyptian scrolls.”

While many have raised concerns about the impact that artificial intelligence could have on human creativity and job loss, those in the archaeological field have begun to see AI as “fundamentally a tool” that can be put to various ends, Nader adds.

In the field of archaeology, AI has already been helpful in locating and identifying numerous finds that were until now inaccessible. One such example is the discovery, using LiDAR laser remote sensing technology, of hidden structures and pyramids constructed by the Maya amid dense Mexican tropical forests. Another includes the identification of a previously unknown L-shaped structure in an ancient Egyptian cemetery in Giza using ground-penetrating technology. In the discipline, AI has proved a helpful tool in myriad other ways as well: The Archaeological Park of Pompeii, for instance, has installed an AI-enabled robot dog to safeguard the site from theft.

“AI is succeeding because of human-shaped data. It’s interesting to see AI play a role at the interface between humanities and these new techniques,” Seales said. “It’s not an accident that you have to have all this human writing to be able to build a large language model. It’s because that human writing captures, in some essence, what it means to be human. That interplay is the next frontier.”